Question

我正在尝试编写一个pyopencl脚本来计算Thomas的循环对称吸引子。的功能是

x'= sin（y）-bx

y'= sin（z）-

z'= sin（x）-bz

我已经用python3编写了一个可以运行的实现，尽管运行缓慢。这是我想要的输出：

working solution

这是我的pyopencl实现的输出：

broken opencl implementation

我相信我在正弦函数上遇到某种舍入误差或近似误差，因此我尝试将所有内容强制转换为双精度，但没有成功。我看到的另一种可能性是，在函数迭代时输出达到的值时出现了一些错误，但是我不知道会是什么。

这里是有问题的内核。

  #pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
    #pragma OPENCL EXTENSION cl_khr_fp64 : enable
    __kernel void thomas(__global float3 *a,
                     __global float3 *output, ulong const maxiter, float const stepSize, float const b )
    {
        int gid = get_global_id(0);

            double x = a[gid][0];
            double y = a[gid][1];
            double z = a[gid][2];
            double x1,y1,z1 = 0.0;
            for(int citer = 0; citer<maxiter;citer++){

                x1 = x+stepSize*(sin(y)-b*x);
                y1 = y+stepSize*(sin(z)-b*y);
                z1 = z+stepSize*(sin(x)-b*z);
                output[gid*maxiter+citer][0]=x1;
                output[gid*maxiter+citer][1]=y1;
                output[gid*maxiter+citer][2]=z1;
                x = x1;
                y = y1;
                z = z1;
            }



    }

a是一个起始值数组，而输出是一个长度为a * maxiter的数组

我希望pyopencl实现的输出与python3实现相匹配，但似乎只在xy平面中输出一个形状，该形状与3d形状的关系对我来说是不确定的。

编辑：这是违规程序的其余代码

import numpy as np
import pyopencl as cl
import open3d as o3d
def calc_thomas_opencl(npoints, stepSize, maxiter, b):
    ballRadius = .5
    ctx = cl.create_some_context()
    queue = cl.CommandQueue(ctx)
    mf = cl.mem_flags
    points = []
    for point in range(npoints):
        x1 = np.random.rand()-.5
        x2 = np.random.rand()-.5
        x3 = np.random.rand()-.5
        u = np.random.rand()
        fac = ballRadius*u**.3/(np.sqrt(x1**2+x2**2+x3**2))
        point = (x1*fac,x2*fac,x3*fac)
        points.append(point)
    a=np.array(points,dtype = np.float32)
    print(a[0])
    a_opencl = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a)
    output = np.zeros([npoints*maxiter,3])
    output_opencl = cl.Buffer(ctx, mf.WRITE_ONLY, output.nbytes)
    prg = cl.Program(ctx, """
    #pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
    #pragma OPENCL EXTENSION cl_khr_fp64 : enable
    __kernel void thomas(__global float3 *a,
                     __global float3 *output, ulong const maxiter, float const stepSize, float const b )
    {
        int gid = get_global_id(0);

            double x = a[gid][0];
            double y = a[gid][1];
            double z = a[gid][2];
            double x1,y1,z1 = 0.0;
            for(int citer = 0; citer<maxiter;citer++){

                x1 = x+stepSize*(sin(y)-b*x);
                y1 = y+stepSize*(sin(z)-b*y);
                z1 = z+stepSize*(sin(x)-b*z);
                output[gid*maxiter+citer][0]=x1;
                output[gid*maxiter+citer][1]=y1;
                output[gid*maxiter+citer][2]=z1;
                x = x1;
                y = y1;
                z = z1;
            }



    }
    """).build()
    prg.thomas(queue, (npoints,), None, a_opencl,
                   output_opencl, np.uint64(maxiter), np.float32(stepSize), np.float32(b))

    cl.enqueue_copy(queue, output, output_opencl).wait()

    return output

xyz = calc_thomas_opencl(1000,.05,1000,.2)
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(xyz)
o3d.visualization.draw_geometries([pcd])

Answer 1

问题出在

 output = np.zeros([npoints*maxiter,3])

它必须是

 output = np.zeros([npoints*maxiter,3], dtype = np.float32)

无法将python3程序转换为pyopencl程序

1 个答案: